Agentic AI Case Studies in Healthcare

Agentic AI case studies in healthcare header naming a pharmacy chain, a health insurer and clinical AI case studies, by Dr. Harish Kotadia, Ph.D.

OpenAI on Monday released computer-use agents for hospitals, insurers and clinicians. With them came four agentic AI case studies in healthcare. The agents click, type and move between screens. So they can work inside record systems and insurer portals that have no link for software. The company’s announcement says a national pharmacy chain and a large health insurer already run them.

The four deployments it describes use the same models but differ on one point. That point is what stands between the agent and the submit button. So the lesson of these agentic AI case studies in healthcare sits in the controls, because the model is the same in every case.

What OpenAI shipped

The release covers the OpenAI API, ChatGPT for Healthcare and the free ChatGPT for Clinicians. A companion post adds better clinical search and more than 20 healthcare plugins.

OpenAI also published test results. Physicians reviewed 325 inbox tasks completed by its GPT-6 Astra model and rated 99 percent of them clinically safe. They called the work very or extremely helpful in 93 percent of 646 assessments.

A cheaper model did almost as well on office work. On HealthAdminBench, a benchmark from Stanford researchers, Astra earned 94.3 percent of the points while GPT-6 Luna earned 91.8 percent. Luna costs 8.5 cents per task, which OpenAI says is 98 percent less.

The controls are named too. Organizations get role-based access and audit logs, and they can add review and approval steps of their own. They can also limit which applications an agent may open.

A backlog cleared in a weekend

A national pharmacy and health services company gives the most concrete account. Its agents send documents, process referrals and support specialist consults across clinical systems, according to the announcement. They finish steps that the systems’ own connections do not fully cover.

One agent cleared a document backlog that had built up for months, and it did so over a single weekend. Another handles the system steps for tens of thousands of consults each day.

The oversight is what I would underline. The company uses validation checks and human review. It also keeps screenshots and action records to confirm what the agent did. It has also published its agent kit as open source, so other teams can repeat the method.

Ten systems, read in seconds

A large health insurer chose a narrower job. Its case managers once had to open more than 10 clinical and administrative systems before calling a member. Now an agent gathers that context in seconds, the announcement says.

The agent reads and assembles, and a person still makes the call. That keeps the risk small. The insurer’s chief information officer said the work went from concept to deployment in under eight weeks, with “the right guardrails for responsible use.”

An agent that stops and asks

A clinical AI company showed a demonstration, not a live rollout. In it, an agent prepares a prior authorization request on an insurer’s website, using evidence from the patient’s chart.

When the record is unclear, the agent stops and asks the clinician. Then it waits for approval before it submits anything. The health system sets which actions the agent may take and can pause it at any moment. It also gets a step-by-step log.

A second clinical AI company is another example. OpenAI describes its inpatient agents, which would clear discharge barriers such as medication access, as planned. Becker’s Hospital Review reports that work as in development.

What the numbers leave out

The safety figure deserves a second look. If 99 percent of 325 tasks were safe, then about three were not, and the announcement does not describe them. The review was also OpenAI’s own.

The test tasks ended in drafts as well. In the published examples, the model saved an unsent reply and an unsigned note for a clinician to approve. So the test measured an agent that proposes, while the pharmacy company’s production agents complete steps.

Outside data shows why that gap matters. An Anthropic study of agent autonomy, published in February, looked at close to a million tool calls on its public API. About 73 percent appeared to have a human involved, and only 0.8 percent of actions appeared irreversible. A submitted authorization request sits in that small second group.

Where I place these cases

I read agentic AI case studies in healthcare through two tools of my own. The first is my Five-Stage Roadmap, which runs from Prompted to Autonomous. The second is the six layers of an agent architecture, where enforcement is the layer teams skip.

On the roadmap, the pharmacy company is Governed for those workflows, because the checks and the evidence are built in. The insurer is Governed too, but on read-only work. The demonstration shows a Governed design that production has not yet proven. And the planned inpatient agents are Piloted at best. Nobody here is Autonomous, and I think that is correct.

Why the controls must move

The layers explain why. A screen is a tool with no schema, so the capability layer just grew to every application a login can reach. But the controls most teams own sit at the API gateway, and a click never passes through it. So enforcement has to move to where the agent works, which is why I say govern the harness, not the model.

Four controls before the first click

For agentic AI case studies in healthcare like these, I would want four things before the first production click.

The demonstration is the closest match, and the pharmacy company’s screenshots are the right evidence. I would still ask each of them one more question. Who can change the approval rule, and does that change get logged? (I ask it on every review, and the answer is usually a pause.)

The click that cannot be undone

I am less sure about the cost story. Luna’s 8.5 cents is a price per task, not per completed and checked task. Review time is the larger number. In regulated loan origination the pattern is the same, because the cheap step is never the one an examiner asks about.

My view is plain. Buyers should judge agentic AI case studies in healthcare by the click that cannot be taken back. So that click is where an agent should earn autonomy last. Instructions in, results out was IT. Intent in, outcomes out is agentic AI, and the outcome includes the record of who approved it.

Which step in your own workflow would you never let an agent submit without a person?

Subscribe to the Blog: I publish a new post every weekday on one agentic AI governance question. Enter your email and each post arrives in your inbox the moment it goes live.

My Books: Three books on agentic AI for regulated enterprises: Intent In, Outcomes Out, Earned Autonomy and The Governed Loop. Reading order, formats and buy links are on one page.

Three books on agentic AI for regulated enterprises by Dr. Harish Kotadia, Ph.D.: Intent In, Outcomes Out; Earned Autonomy; The Governed Loop, with their mybook.to buy links on a Thanksgiving background

© Dr. Harish Kotadia, Ph.D., All Rights Reserved, 2026

Dr. Harish Kotadia, Ph.D., is an Enterprise AI Architect with 20+ years of IT consulting experience serving Fortune 100 clients, specializing in agentic AI systems built on Anthropic Claude, AWS Bedrock, and Google Vertex AI.

Disclaimer: This blog post is based on publicly available academic publications, vendor documentation, open standards, and news items from reputed media sources linked above. This post is intended for educational purposes, to help the enterprise agentic AI community build a shared vocabulary from public, authoritative sources.

Views and opinions expressed here are my own and do not represent those of any employer or client, past or present. The analysis presented is my independent interpretation of the published sources linked above and does not constitute legal, financial, or consulting advice of any kind.

 


Discover more from Agentic AI Governance | Dr. Harish Kotadia, Ph.D.

Subscribe to get the latest posts sent to your email.

Leave a Reply

Discover more from Agentic AI Governance | Dr. Harish Kotadia, Ph.D.

Subscribe now to keep reading and get access to the full archive.

Continue reading